Curve fitting

(for II BA/BSc Statistics Students)

Department of Statistics
P.R. Government College(A)
Kakinada
Curve Fitting
Curve fitting means finding a functional
relationship between the two variables X and
Y in the form Y = f(X) (X-Independent, Y-
Dependent) using a given set of n points
(x1,y1), (x2,y2),....(xn,yn).
The functional form is to be fitted at
best level using a principle, which is called
principle of least squares.
Principle of Least Squares:
The principle of least squares was
introduced by Legendre.
The Principle is “ the sum of squares of
deviations between actual values and estimated
values by the curve is to be minimum”.
i.e., ∑[yi-f(xi)]2 is to be minimum.
-> the term yi-f(xi)=E(say) is called error or residue
of ith ordinate.
Now the objective of the curve fitting is to
estimate the parameters in the functional relation
that chosen by minimizing the total error
 some important curves applied in
practical life are,
-> Straight Line (Y = a + bX) [1st degree polynomial]
-> Parabola (Y=a + bX + CX2) [2 degree polynomial]
nd

-> Exponential Curves (Y=aebX, Y=abX)

-> Power Curve (Y=aXb)
Let us fit these curves for a given data set
using the principle of least squares.
This means to find the best estimates for
the parameters a, b, etc, by applying the
principle.
a). Fitting of a Straight Line:
Let Y = a + bX ...(1) be the straight line to be
fitted for a given data of n points (x1,y1), (x2,y2),...
(xn,yn) using the principle of least squares.
According to the principle,
E = ∑[yi-f(xi)]2 is to be minimum.
i.e., E =∑[yi-(a+bxi)]2 is to be minimum in the
variations of a and b.
To minimize E, we have to
solve ∂E/∂a =0 and ∂E/∂b =0
and prove ∂2E/∂a2 > 0 and ∂2E/∂b2 > 0.
Consider, ∂E/∂a = 0 Now consider, ∂E/∂b= 0
=> ∂(∑[yi-(a+bxi)]2 )/∂a = 0 => ∂(∑[yi-(a+bxi)]2 )/∂b = 0

=> 2 {∑[yi-(a+bxi)](-1)} = 0 => 2 {∑[yi-(a+bxi)](-xi)} = 0

=> ∑yi – a ∑(1) - b∑xi = 0 => ∑xiyi – a ∑xi - b∑xi2 = 0
=> ∑yi = na + b∑xi ....(2) => ∑xiyi = a∑xi + b∑xi2 ...(3)
The equations (2) and (3) are called Normal equations.
Solving these equations we can get best estimates for a
and b.
With these estimates line (1) becomes the line of best
fit.
-> we can prove easily ∂2E/∂a2 > 0 and ∂2E/∂b2 > 0 at
that values of a and b (Try)
b). Fitting of a Parabola:
Let Y = a + bX + cX2 ...(1) be the parabola to be
fitted for a given data of n points (x1,y1), (x2,y2),...
(xn,yn) using the principle of least squares.
According to the principle,
E = ∑[yi-f(xi)]2 is to be minimum.
i.e., E =∑[yi-(a+bxi+cxi2)]2 is to be minimum in the
variations of a, b and c.
To minimize E, we have to
solve ∂E/∂a =0, ∂E/∂b =0 and ∂E/∂c =0
and prove ∂2E/∂a2 > 0, ∂2E/∂b2 > 0 and ∂2E/∂c2 > 0
Consider, ∂E/∂a = 0
=> ∂(∑[yi-(a+bxi+cxi2)]2 )/∂a = 0

=> 2 {∑[yi-(a+bxi+cxi2)](-1)} = 0
=> ∑yi – a ∑(1) - b∑xi - c∑xi2 = 0
=> ∑yi = na + b∑xi + c∑xi2 ....(2)
Again consider, ∂E/∂b= 0
=> ∂(∑[yi-(a+bxi+cxi2)]2 )/∂b = 0

=> 2 {∑[yi-(a+bxi+cxi2)](-xi)} = 0
=> ∑xiyi – a ∑xi - b∑xi2 - c∑xi3 = 0
2 3
Now, ∂E/∂c = 0
=> ∂(∑[yi-(a+bxi+cxi2)]2 )/∂a = 0

=> 2 {∑[yi-(a+bxi+cxi2)](-xi2)} = 0
=> ∑xi2yi – a ∑xi2 - b∑xi3 - c∑xi4 = 0
=> ∑xi2yi = a ∑xi2 + b∑xi3 + c∑xi4 ....(4)
The equations (2), (3) and (4) are called Normal
equations. Solving these equations we can get best
estimates for a and b.
With these estimates (1) becomes the parabola of best
fit.
c). Fitting of an Exponential curve of the form
Y = aebX:
Y = aebX ...(1) is the exponential curve to be fitted
for a given data of n points (x1,y1), (x2,y2),...(xn,yn) using
the principle of least squares.
If we approach directly, we can not get normal equations
( because of exponential terms).
So, we can simplify the work using logarithms.
Taking Ln (loge)on both sides of (1)
Ln(Y) = Ln(aebX)
=> Ln(Y) = Ln(a) + bX ....(2)
Clear that (2) is a straight line in Ln(Y) and X with
parameters Ln(a) and b.
By applying the principle of least squares, the Normal
equations to estimate the parameters Ln(a) and b are,

∑Ln(Y) = nLn(a) + b∑X

and ∑XLn(Y) = Ln(a)∑X + b∑X2

we can get ‘a’ as , a = eLn(a)

with a and b estimates, the curve (1) is an exponential
curve of best fit.
d). Fitting of an Exponential curve of the form
Y = abX:
Y = abX ...(1) is the exponential curve to be fitted
for a given data of n points (x1,y1), (x2,y2),...(xn,yn) using
the principle of least squares.
If we approach directly, we can not get normal equations
( because of exponential terms).
So, we can simplify the work using logarithms.
Taking Ln (i.e., loge)on both sides of (1)
Ln(Y) = Ln(abX)
=> Ln(Y) = Ln(a) + Ln(b)X ....(2)
Clear that (2) is a straight line in Ln(Y) and X with
parameters Ln(a) and Ln(b).
By applying the principle of least squares, the Normal
equations to estimate the parameters Ln(a) and Ln(b)
are,

∑Ln(Y) = nLn(a) + Ln(b)∑X

and ∑XLn(Y) = Ln(a)∑X + Ln(b)∑X2

we can get ‘a’ and ‘b’ as , a = eLn(a) and b= eLn(b)

with a and b estimates, the curve (1) is an exponential
curve of best fit.
e). Fitting of a power curve of the form
Y = aXb:
Y = aXb ...(1) is the power curve to be fitted for a
given data of n points (x1,y1), (x2,y2),...(xn,yn) using the
principle of least squares.
If we approach directly, we can not get normal equations
( because of exponential terms).
So, we can simplify the work using logarithms.
Taking Ln (i.e., loge)on both sides of (1)
Ln(Y) = Ln(aXb)
=> Ln(Y) = Ln(a) + bLn(X) ....(2)
Clear that (2) is a straight line in Ln(Y) and Ln(X) with
parameters Ln(a) and b.
By applying the principle of least squares, the Normal
equations to estimate the parameters Ln(a) and b are,

∑Ln(Y) = nLn(a) + b∑Ln(X)

and ∑Ln(X)Ln(Y) = Ln(a)∑Ln(X) + b∑[Ln(X)]2

we can get ‘a’ as , a = eLn(a)

with a and b estimates, the curve (1) is a power curve
of best fit.
Uses of Curve fitting:
1. It is useful for extrapolation effectively and also for
interpolation.
2. It is useful for future forecasting, which is essential for
business people.
3. It is useful for studying the relationship between the two
variables.
4. It is useful for studying regression analysis and time
series analysis.
Note: The word “best fit” depends on the choice of an appropriate curve
for a given data. It can be done as given below.
-> the 1st order differences of Y, ∆Y ~0  Straight Line
-> the 1st order differences of Y, ∆2Y ~0  Parabola
->the 1st order differences of Ln(Y),∆Ln(Y)~0Exponential curves
Problem:
1. Fit a straight line to the following data by the method of
leastX squares.
0 1 2 3 4 5
Y 5 13 25 36 48 60

Solution: Let the straight line, Y =a + bX ....(1)

By the principle of least squares, the normal
X equations
Y XYto estimate
X2
a and b are, 0 5 0 0
∑Y = na + b∑X ----(2) 1 13 13 1
∑XY = a∑X + b∑X2 ....(3) 2 25 50 4
3 36 108 9
Substituting the totals and n=6
4 48 192 16
in (2)&(3) 5 60 300 25
187 = 6a + 15b ....(4) 15 187 663 55
663 = 15a + 55b ....(5)
Solving (4)&(5),
a = 3.2381, b = 11.1714
Problem:
2. Fit a parabola to the following data by the method of least squares.
X 0 1 2 3 4
Y 1 6 17 34 57

Solution: Let the parabola, Y =a + bX + cX2 ....(1)

By the principle of least squares, the normal equations to 2estimate a, b 3and 4
X Y XY X X 2Y X X
c are,
∑Y = na + b∑X + c∑X2 ----(2) 0 1 0 0 0 0 0
∑XY = a∑X + b∑X2+ c∑X3....(3) 1 6 6 1 6 1 1
∑X2Y = a∑X2 + b∑X3 + c∑X4 ---(4)
2 17 34 4 68 8 16
Substituting the totals and n=5
3 34 102 9 306 27 81
in (2),(3)&(4)
115 = 5a + 10b + 30c ....(5) 4 57 228 16 912 64 256
370 = 10a + 30b + 100c ....(6) 10 115 370 30 1292 100 354
1292 = 30a + 100b + 354c ---(7)
Solving (5),(6)&(7)
a = 1, b = 2 , c = 3
Problem:
3.Fit an exponential curve of the form Y=abX to the following data by the
method of least squares.
X 1 2 3 4 5 6 7 8
Y 1 1.2 1.8 2.5 3.6 4.7 6.6 9.1

Solution: Given , Y =a bX ....(1) => Ln(Y) = Ln(a) + Ln(b) X ...(2)

By the principle of least squares, the normal equations to estimate Ln(a) and
Ln(b) in (2)are, X Y Ln(Y) XLn(Y) X2
∑Ln(Y) = nLn(a) + Ln(b)∑X ----(3) 1 1 0 0 1
∑XLn(Y) = Ln(a)∑X + Ln(b)∑X2....(4) 2 1.2 0.1823 0.3646 4
Substituting the totals and n=8 3 1.8 0.5878 1.7634 9
in (3)&(4) 4 2.5 0.9163 3.6652 16
8.6103 = 8Ln(a) + 36Ln(b) ....(5) 5 3.6 1.2809 6.4045 25
52.3594 = 36Ln(a) + 204Ln(b)..(6) 6 4.7 1.5476 9.2856 36
Solving (5)&(6), 7 6.6 1.8871 13.2097 49
Ln(a) = -0.3823, Ln(b) = 0.3241 8 9.1 2.2083 17.6664 64
=> a = 0.6823 , b = 1.3828 36 ** 8.6103 52.3594 204
From (1), the required curve is,
Self Assessment Questions:
1. What is curve fitting? Explain.
2. Explain the principal of least squares.
3. Explain the fitting of straight line using the
method of least squares
4. Explain the fitting of second degree parabola.
5. Explain the fitting exponential curves.
6. Explain the fitting of power curve.
7. Fit a straight line to the data given data by the
method of least squares.
X 10 20 30 40 50 60 70
Y 32 62 92 122 152 182 212